Chinese AI companies have begun using optimisation tools to compensate for their limited access to Nvidia chips for demanding inference tasks, such as programming, a local company said.
Approaching.AI, a start-up that specialises in inference optimisation software, said its revenue in June surpassed its total revenue for all of 2025, amid rising demand for such technology.
Approaching.AI said its inference engine, KTransformers, has been used to run DeepSeek’s full-sized model on a single Nvidia RTX 4090 GPU, combined with a CPU and large amounts of memory, demonstrating the efficiencies that are possible with such an approach.
Beijing, China. Image credit: Henry Chen/Unsplash
Heterogeneous inference
